Anthropic expands partnership with Google and Broadcom for multiple gigawatts of next-generation compute

Anthropic signed a deal with Google and Broadcom for multiple gigawatts of next‑generation TPU capacity expected to come online starting in 2027 to power its Claude models. Run‑rate revenue has surpassed $30 billion (up from ~$9 billion at end‑2025) and the number of enterprise customers spending >$1M annually has doubled to over 1,000, underscoring rapid commercial traction; the bulk of new capacity will be sited in the U.S., building on a prior $50 billion U.S. compute commitment.

Analysis

The multi-vendor compute path favours platform owners that can monetize differentiated software and operational SLAs rather than raw silicon sales; that structurally benefits cloud franchises with sticky enterprise contracts and upsell levers (service tiers, managed inference, security). Secondary beneficiaries will be suppliers of high-throughput networking, custom firmware and systems integration where margins are less cyclic and lead times create pricing power; this shifts value capture away from commodity GPU spot markets into bespoke infrastructure stacks over a multi-year window.

Key risks cluster around demand concentration, policy interventions, and hardware obsolescence. Heavy enterprise reliance on a single advanced LLM supplier or on proprietary accelerator stacks can reverse quickly if open-source models narrow the performance gap or if regulation forces on‑premise alternatives; vendor leverage can flip to customers in 6–18 months when procurement cycles and TCO analyses are re-run. Supply-side shocks — constrained ASIC/NIC capacity or Broadcom-style pricing power — can amplify margins short-term but also prompt customers to accelerate multi-cloud or in-house silicon plans.

Consensus is underweighting the dampening effect of platform diversification on pure-play GPU vendors while over-counting the near-term durable margins for any single cloud provider. The more sensible positioning is to favor firms exposed to non-compute adjuncts (networking, custom silicon IP, integration services) and cloud operators that can wrap higher-margin software around infrastructure, while treating commodity GPU exposure as a timing-sensitive trade rather than a permanent winner.

Key Decisions for Investors

• Long AVGO (12–24 months): buy a modest position or a 12‑month call spread sized 3–5% portfolio risk to capture higher ASPs for high-speed interconnects and custom ASIC content; target 30–50% upside if enterprise buildouts continue, stop-loss at 12% to limit drawdown.
• Long GOOGL (9–18 months, options collar): initiate a 6–9 month out call-buy with a protective put (costed collar) to play cloud software monetization and TPU-adjacent services; asymmetric risk/reward 1:3 assuming continued enterprise AI spend, take profits if shares outperform by 25–30%.
• Pair trade (6–12 months): long AMZN (AWS exposure to managed AI services) / short NVDA (GPU training share risk) in equal dollar notional to express preference for platform + services over commodity training hardware; size to 2–4% net portfolio risk, expect mean reversion if training demand diversifies, unwind if NVDA reports stronger-than-expected pricing power or new product cycle.
• Event-driven alpha: buy AVGO or GOOGL on pullbacks >5% tied to hardware supply disruption headlines, trim into strength; conversely short names tightly levered to commodity GPU pricing on positive headlines about open-source LLMs reducing training spend (stop at 8% adverse move).